Statistically robust linear and nonlinear wavelet analysis applied to plasma edge turbulence

nchez, M. A. Pedrosa, R. Balb n, B. Ph. van Milligen, C. Hidalgo, E. Sa a-Corte s and I. Garc
Asociacion Euratom-CIEMAT, Madrid, Spain

G. R. Tynan
Institute for Plasma and Fusion Research, University of California, Los Angeles, California 90024

Presented on 15 May 1996 The analysis of turbulence requires adequate analysis tools. In particular, two general properties of turbulence make the application of standard analysis tools spectral analysis difcult: 1 intermittency, which causes the characteristics of turbulence to change on a relatively short time scale, implies that analysis tools must not integrate over time scales longer than the intermittent time scale; 2 nonlinearity, a basic property of all numerical models of chaos and turbulence, requires specially adapted tools for its proper detection. In the present work, we develop a statistically stable formalism for the application of wavelet analysis techniques to the analysis of turbulence. We apply these techniques to the analysis of Langmuir probe data in the plasma edge region of the TJ-IU Torsatron. We draw some tentative conclusions about radial correlation lengths. Further, we analyze the L/H transition at the Continous Current Tokamak and reconstruct a radial prole of bicoherence in the plasma edge in the H-mode. 1997 American Institute of Physics. S0034-67489765601-4

I. INTRODUCTION

II. SPECTRAL ANALYSIS TOOLS

Although the phenomenon of turbulence is only partially understood, there seems to be consensus on several aspects. First, that intermittency is a basic property of turbulence. This means that the characteristics of the turbulence spectral distribution, amplitude, etc. vary on a short time scale. Analysis techniques that rely on the accumulation of data over time scales larger than this characteristic time scale will then average out much of the dynamics and obliterate relevant information as may occur with Fourier analyses. Wavelet analysis is ideally suited to tackle this problem. Standard wavelet analysis is, however, not fully adequate since it is not statistically robust. In this article, we shall develop statistically stable wavelet techniques. Second, it is generally accepted that turbulence only arises in nonlinear systems. Therefore, to understand the nature of turbulence, it is essential to employ analysis tools that are capable of handling this nonlinearity. The usual analyses based on cross- spectra and cross- correlations, essentially linear analysis techniques, are not adequate. Nonlinear analysis tools can be obtained by generalizing the common spectral analysis methods to higher order, which then are sensitive to nonlinear interactions. We focus on the so-called bispectral analysis, a method for the detection of quadratic interactions. Statistical stability of the bispectruma third-order spectrumis again an important point, and noise level estimates can be provided. The main application of the bispectrum is the detection of phase coupling. It has been shown in earlier publications1,2 how reasonable time resolution, relevant to intermittency in some turbulent phenomena, can be achieved. Also it has been discussed how the bicoherence is related to nonlinearity and the presence of structure in turbulence.
Rev. Sci. Instrum. 68 (1), January 1997

In this section, we give a statistically robust denition of the wavelet cross spectrum and cross coherence.
A. Wavelet analysis

In the present work, we use the continuous wavelet transform based on the Morlet wavelet which has the benet of conceptual closeness to the Fourier analysis base functions e i t :
1 2 t 1/4 exp i 2 t 2 t ;

a t a 1/2 t / a .

We assign a frequency 2/a to each scale a . The frequency resolution of the wavelet a ( t ) is /4, and the time resolution is t 2 a . Note that t , independent of a . The wavelet transform of a function f ( t ) is dened by3,4 W f a,

f t a t dt .

The advantage of the wavelet analysis lies in the fact that the time resolution is variable with frequency, so that high frequencies have a sharper time resolution. Thus, turbulent signals are decomposed into oscillations that die out in time, and more rapidly so the higher their frequency, which seems more natural for turbulence than the Fourier picture of stationary oscillations.
B. Statistically robust wavelet spectral analysis

In order to obtain statistical stability for our spectral analysis tools while maintaining time resolution, we average appropriate combinations of wavelet coefcients over a small nite time interval T : T 0 T /2 T 0 T /2 . This procedure allows the estimation of a noise level which will tell us the statistical signicance of the obtained results.
1997 American Institute of Physics 967

0034-6748/97/68(1)/967/4/$10.00

Thus, e.g., the wavelet cross spectrum is given by Cw f g a , T 0

W* f a , W g a , d ,

where f ( t ) and g ( t ) are two time series and C w f g is complex. The normalized wavelet cross coherence is

w f g a , T 0

TW * f a , W g a , d
w 1/2 Pw f a , T 0 P g a , T 0

which can take on values between 0 and 1. Here the wavelet w auto-power spectrum is P w f ( a , T 0 ) C f f ( a , T 0 ). w To calculate f g Eq. 4, the wavelet coefcients are determined for each of N T f samp samples in the interval T and averaged Eq. 2, where samp2 f samp is the sampling frequency. However, these wavelet coefcients are not all statistically independent, since the chosen wavelet family is not orthogonal. Due to the periodicity a of the wavelets, two statistically independent estimates of the wavelet coefcients are separated by a time a /2, or by a number of points M ( a ) a samp/4. Thus, the integral appearing in Eq. 2 is carried out over N / M ( a ) independent estimates of wavelet coefcients. Using this, we may estimate the statistical noise level as

w f g a , T 0 2

samp 1 N

1/2

The value of w f g as obtained for Gaussian noise conforms quite well to this prediction. Similar denitions can be given for higher-order spectra that measure the presence of nonlinear interactions. Refer to Refs. 1 and 2 for the corresponding denitions.

III. RADIAL COHERENCE OBSERVED ON THE TJ-IU TORSATRON

FIG. 1. a Cross spectrum and cross coherence of I sat from two radially separated probes in the edge zone of a TJ-IU discharge. Continuous line: average wavelet spectrum; long dashes: wavelet coherence; short dashes: the noise level of the coherence. b Same for the instantaneous particle ux derived from the probe data see the text. The value of the coherence is smaller than for I sat .

/B , E The time-resolved radial turbulent ux, T n T was measured at two radially separated positions separation 1 cm in the plasma edge region of the TJ-IU Torsatron.5 The plasmas were heated by electron cyclotron resonance heating ECRH, and have R 0.6 m, central rotational transform 00.21, a 0.1 m, B t 0.6 T, and n e 51018 m3. Each of the two probes has three tips, aligned perpendicular to the magnetic eld and separated poloidally by 0.2 cm. The probes were designed and positioned to avoid shadowing effects.6 The uctuating poloidal electric eld is then derived from the measured uctuating oating potential in (1) (2)]/ . [ the two most extreme probe tips: E f f The third probe tip is located between the other two, but slightly displaced toroidally. It is set up to measure the ion saturation current I sat . The turbulent ux T has been calculated neglecting the inuence of temperature uctuations n I sat. The signals are sampled at 1 MHz. Figure 1a shows the cross spectrum and radial cross correlation between the I sat signals of the two radially separated probes. The inuence of a magnetohydrodynamics MHD mode is recognized in the peak of the spectrum at about 20 kHz. The cross correlation peaks at slightly higher frequency 35 kHz.
968 Rev. Sci. Instrum., Vol. 68, No. 1, January 1997

Figure 1b shows the same graph, now calculated for T . Features similar to the ones observed in Fig. 1a can be seen, although at higher frequencies. This frequency shift is easily explained by the fact that T is a quadratic signal. What is most interesting is that the ux does not show a either, not shown. larger radial correlation than I sat or f Figure 2 shows the temporally resolved coherence of I sat . The time resolution is 0.5 ms. The noise level is the same as in Fig. 1. The coherence is highly intermittent and occasionally very high values are achieved much higher than the time-average value.
IV. BICOHERENCE PROFILE AT THE L/H TRANSITION ON CCT

The data analyzed in this section are from the continuous current tokamak CCT. It was operated with R 1.5 m, a 0.35 m, B t 0.25 T, I p 40 kA, n e 21018 m3, V loop1.21.4 V, T e 0150 eV, and T i 0100 eV. H modes were induced by biasing a small electrode located about 0.1 m inside the limiter radius with respect to the vessel wall.7
Plasma diagnostics

FIG. 3. Analysis of the saturation current I sat measured by a Langmuir probe positioned on the outside midplane at CCT. Long dashes: rms of I sat . Drawn line: bicoherence of I sat . Short dashes: Noise level of the bicoherence. The gray area indicates the H -mode period, induced by probe biasing. FIG. 2. Time-resolved wavelet coherence-versus-time graph of I sat . The coherence is between two radially separated probes see the text. The noise level is the same as in Fig. 1. In several time intervals and at various frequencies the coherence obtains values far above the time-average value.

For studies of edge turbulence and transport, a poloidal Langmuir probe array was used.8,9 It was congured such that one of the probe tips was recording the oating potential locally. The sampling rate was 2.5 MHz. Here we focus on data from a Langmuir probe located at the outside midplane. Prior to the initiation of H mode, the probe was located just outside the last-closed-ux-surface LCFS. During H modes, the increasing plasma pressure causes a slow movement of the plasma column out towards the low-eld side. Thus, when the H mode electrode bias is suddenly turned on, the outside midplane probe records a slow increase in the negative dc oating potential. This enabled us to reconstruct the radial prole in the H mode of the quantities measured by the probe using an estimate of the instantaneous probe position relative to the LCFS. Knowing the value of the radial electric eld E r , the radial position r of the probe can be estimated as r f / E r , where f is the oating potential measured by the probe. From Doppler shift measurements, we estimated the electric eld to be E r 100 V/cm. Uncertainties in this estimate translate into an uncertainty of 30% 50% in the absolute reconstructed position, although the relative position is much more accurate since E r does not vary signicantly during the measurements. Figure 3 shows the rms values and the bicoherence of the measured ion saturation current I sat for the outside midplane probe. The H -mode period gray area shows a slight reduction of the rms and a gradual increase of the bicoherence as the plasma moves outward. A broad range of frequencies is involved in the production of the high bicoherence around t 80 ms, with predominance of frequencies around 250 and 500 kHz. Observations reported earlier for L / H transitions using reectometry2 showed an abrupt increase of the bicoherence and decrease of the rms at the transition. The difference with the present apparently smooth transition may be explained by the fact that here the probe location is initially at the LCFS and moves gradually inward, whereas in the measureRev. Sci. Instrum., Vol. 68, No. 1, January 1997

ments reported earlier the measurements were taken well inside the LCFS. This would imply that the gradual change observed is due rather to the existence of a bicoherence prole than to a slow temporal change. Performing standard statistical analysis on the outside midplane I sat signal, we calculate the probability distribution function PDF before and after the transition. Calculations are performed on records of 12 500 samples after high-pass digital ltering with a cutoff frequency 1 kHz to remove drifts. We nd that the L -mode PDF is Poisson-like with skewness S 1, and kurtosis K 5, whereas in the H mode it is more like a Gaussian 0S 1, 3K 4 Fig. 4. The deviation from gaussianity in the L mode as contrasted with the near gaussianity in the H mode is consistent with earlier studies of the relation between the PDF shape of turbulent signals and plasma conditions.10 Using the above-mentioned estimate of the probe position, the prole of bicoherence was reconstructed during the H -mode phase from the signal of the outside midplane probe. The result is shown in Fig. 5, along with a similar prole obtained from a different but similar discharge. We recall that a high value of the bicoherence may either indicate the presence of nonlinear interactions or of quasistatic structure.11 It is interesting that these features should occur a

FIG. 4. PDF of the outside midplane I sat signal in the L mode and the H mode. L mode 6570 ms: Skewness S 1.320.07, Kurtosis K 5.94 0.22; H mode 7277 ms: S 0.150.06, K 3.270.11; H mode 8590 ms: S 0.670.06, K 3.830.13. Plasma diagnostics 969

ture, reported for the rst time here, is as yet unclear but may have bearing on H -mode physics.
nchez, Phys. Rev. Lett. 74, 395 B. Ph. van Milligen, C. Hidalgo, and E. Sa 1995. 2 nchez, T. Estrada, C. Hidalgo, B. Bran as, B. B. Ph. van Milligen, E. Sa a, Phys. Plasmas 2, 3017 1995. Carreras, and L. Garc 3 C. Chui, An Introduction to Wavelets Academic, New York, 1992. 4 I. Daubechies, Ten Lectures on Wavelets, National Science Foundation Series In Applied Mathematics, S.I.A.M. 1992 unpublished. 5 bar et al. Plasma Phys. Control. Nucl. Fusion Research, IAEAE. Ascas CN-60/A6-1, 1, 749 1994. 6 nchez, R. Balb n, I. M. A. Pedrosa, C. Hidalgo, B. van Milligen, E. Sa a-Corte s, H. Niedermeyer, and L. Giannone, Proceedings of the 23rd Garc European Conference Kiev, 1996 unpublished. 7 R. J. Taylor, M. L. Brown, B. D. Fried, H. Grote, J. R. Liberati, G. J. Morales, P. Pribyl, D. Darrow, and M. Ono, Phys. Rev. Lett. 63, 2365 1989. 8 G. R. Tynan, Ph.D. thesis, School of Engineering, University of California, Los Angeles, 1991. 9 G. R. Tynan, L. Schmitz, R. W. Conn, R. Doerner, and R. Lehmer, Phys. Rev. Lett. 68, 3032 1992. 10 nchez, M. A. Pedrosa, R. Balb n, I. B. A. Carreras, C. Hidalgo, E. Sa a-Corte s, B. Ph. van Milligen, D. E. Newman, and V. E. Lynch, Garc Phys. Plasmas 3, 2664 1996. 11 cseli and J. Trulsen, Plasma Phys. Control. Fusion 35, 1701 H. L. Pe 1993.
1

FIG. 5. Bicoherence prole in the H mode deduced from the outside midplane probe signal shown in Fig. 3 and an estimate of the probe position relative to the LCFS continuous line. Also included is a similar prole for another, similar discharge long dashes. The noise level of the bicoherence is indicated by the short-dashed line.

small distance inside and not at the LCFS, which is possibly related to a decorrelating velocity shear layer at the LCFS. The results also seem to indicate a maximum of bicoherence about half a cm inside the LCFS. The meaning of this fea-

970

Rev. Sci. Instrum., Vol. 68, No. 1, January 1997

Plasma diagnostics